1 /* 2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com> 3 * 4 * Parts came from builtin-{top,stat,record}.c, see those files for further 5 * copyright notes. 6 * 7 * Released under the GPL v2. (and only v2, not any later version) 8 */ 9 #include <poll.h> 10 #include "cpumap.h" 11 #include "thread_map.h" 12 #include "evlist.h" 13 #include "evsel.h" 14 #include "util.h" 15 #include "debug.h" 16 17 #include <sys/mman.h> 18 19 #include <linux/bitops.h> 20 #include <linux/hash.h> 21 22 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y)) 23 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y) 24 25 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus, 26 struct thread_map *threads) 27 { 28 int i; 29 30 for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i) 31 INIT_HLIST_HEAD(&evlist->heads[i]); 32 INIT_LIST_HEAD(&evlist->entries); 33 perf_evlist__set_maps(evlist, cpus, threads); 34 } 35 36 struct perf_evlist *perf_evlist__new(struct cpu_map *cpus, 37 struct thread_map *threads) 38 { 39 struct perf_evlist *evlist = zalloc(sizeof(*evlist)); 40 41 if (evlist != NULL) 42 perf_evlist__init(evlist, cpus, threads); 43 44 return evlist; 45 } 46 47 static void perf_evlist__purge(struct perf_evlist *evlist) 48 { 49 struct perf_evsel *pos, *n; 50 51 list_for_each_entry_safe(pos, n, &evlist->entries, node) { 52 list_del_init(&pos->node); 53 perf_evsel__delete(pos); 54 } 55 56 evlist->nr_entries = 0; 57 } 58 59 void perf_evlist__exit(struct perf_evlist *evlist) 60 { 61 free(evlist->mmap); 62 free(evlist->pollfd); 63 evlist->mmap = NULL; 64 evlist->pollfd = NULL; 65 } 66 67 void perf_evlist__delete(struct perf_evlist *evlist) 68 { 69 perf_evlist__purge(evlist); 70 perf_evlist__exit(evlist); 71 free(evlist); 72 } 73 74 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry) 75 { 76 list_add_tail(&entry->node, &evlist->entries); 77 ++evlist->nr_entries; 78 } 79 80 int perf_evlist__add_default(struct perf_evlist *evlist) 81 { 82 struct perf_event_attr attr = { 83 .type = PERF_TYPE_HARDWARE, 84 .config = PERF_COUNT_HW_CPU_CYCLES, 85 }; 86 struct perf_evsel *evsel = perf_evsel__new(&attr, 0); 87 88 if (evsel == NULL) 89 return -ENOMEM; 90 91 perf_evlist__add(evlist, evsel); 92 return 0; 93 } 94 95 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist) 96 { 97 int nfds = evlist->cpus->nr * evlist->threads->nr * evlist->nr_entries; 98 evlist->pollfd = malloc(sizeof(struct pollfd) * nfds); 99 return evlist->pollfd != NULL ? 0 : -ENOMEM; 100 } 101 102 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd) 103 { 104 fcntl(fd, F_SETFL, O_NONBLOCK); 105 evlist->pollfd[evlist->nr_fds].fd = fd; 106 evlist->pollfd[evlist->nr_fds].events = POLLIN; 107 evlist->nr_fds++; 108 } 109 110 static void perf_evlist__id_hash(struct perf_evlist *evlist, 111 struct perf_evsel *evsel, 112 int cpu, int thread, u64 id) 113 { 114 int hash; 115 struct perf_sample_id *sid = SID(evsel, cpu, thread); 116 117 sid->id = id; 118 sid->evsel = evsel; 119 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS); 120 hlist_add_head(&sid->node, &evlist->heads[hash]); 121 } 122 123 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel, 124 int cpu, int thread, u64 id) 125 { 126 perf_evlist__id_hash(evlist, evsel, cpu, thread, id); 127 evsel->id[evsel->ids++] = id; 128 } 129 130 static int perf_evlist__id_add_fd(struct perf_evlist *evlist, 131 struct perf_evsel *evsel, 132 int cpu, int thread, int fd) 133 { 134 u64 read_data[4] = { 0, }; 135 int id_idx = 1; /* The first entry is the counter value */ 136 137 if (!(evsel->attr.read_format & PERF_FORMAT_ID) || 138 read(fd, &read_data, sizeof(read_data)) == -1) 139 return -1; 140 141 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 142 ++id_idx; 143 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 144 ++id_idx; 145 146 perf_evlist__id_add(evlist, evsel, cpu, thread, read_data[id_idx]); 147 return 0; 148 } 149 150 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id) 151 { 152 struct hlist_head *head; 153 struct hlist_node *pos; 154 struct perf_sample_id *sid; 155 int hash; 156 157 if (evlist->nr_entries == 1) 158 return list_entry(evlist->entries.next, struct perf_evsel, node); 159 160 hash = hash_64(id, PERF_EVLIST__HLIST_BITS); 161 head = &evlist->heads[hash]; 162 163 hlist_for_each_entry(sid, pos, head, node) 164 if (sid->id == id) 165 return sid->evsel; 166 return NULL; 167 } 168 169 union perf_event *perf_evlist__read_on_cpu(struct perf_evlist *evlist, int cpu) 170 { 171 /* XXX Move this to perf.c, making it generally available */ 172 unsigned int page_size = sysconf(_SC_PAGE_SIZE); 173 struct perf_mmap *md = &evlist->mmap[cpu]; 174 unsigned int head = perf_mmap__read_head(md); 175 unsigned int old = md->prev; 176 unsigned char *data = md->base + page_size; 177 union perf_event *event = NULL; 178 179 if (evlist->overwrite) { 180 /* 181 * If we're further behind than half the buffer, there's a chance 182 * the writer will bite our tail and mess up the samples under us. 183 * 184 * If we somehow ended up ahead of the head, we got messed up. 185 * 186 * In either case, truncate and restart at head. 187 */ 188 int diff = head - old; 189 if (diff > md->mask / 2 || diff < 0) { 190 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n"); 191 192 /* 193 * head points to a known good entry, start there. 194 */ 195 old = head; 196 } 197 } 198 199 if (old != head) { 200 size_t size; 201 202 event = (union perf_event *)&data[old & md->mask]; 203 size = event->header.size; 204 205 /* 206 * Event straddles the mmap boundary -- header should always 207 * be inside due to u64 alignment of output. 208 */ 209 if ((old & md->mask) + size != ((old + size) & md->mask)) { 210 unsigned int offset = old; 211 unsigned int len = min(sizeof(*event), size), cpy; 212 void *dst = &evlist->event_copy; 213 214 do { 215 cpy = min(md->mask + 1 - (offset & md->mask), len); 216 memcpy(dst, &data[offset & md->mask], cpy); 217 offset += cpy; 218 dst += cpy; 219 len -= cpy; 220 } while (len); 221 222 event = &evlist->event_copy; 223 } 224 225 old += size; 226 } 227 228 md->prev = old; 229 230 if (!evlist->overwrite) 231 perf_mmap__write_tail(md, old); 232 233 return event; 234 } 235 236 void perf_evlist__munmap(struct perf_evlist *evlist) 237 { 238 int cpu; 239 240 for (cpu = 0; cpu < evlist->cpus->nr; cpu++) { 241 if (evlist->mmap[cpu].base != NULL) { 242 munmap(evlist->mmap[cpu].base, evlist->mmap_len); 243 evlist->mmap[cpu].base = NULL; 244 } 245 } 246 } 247 248 int perf_evlist__alloc_mmap(struct perf_evlist *evlist) 249 { 250 evlist->mmap = zalloc(evlist->cpus->nr * sizeof(struct perf_mmap)); 251 return evlist->mmap != NULL ? 0 : -ENOMEM; 252 } 253 254 static int __perf_evlist__mmap(struct perf_evlist *evlist, struct perf_evsel *evsel, 255 int cpu, int prot, int mask, int fd) 256 { 257 evlist->mmap[cpu].prev = 0; 258 evlist->mmap[cpu].mask = mask; 259 evlist->mmap[cpu].base = mmap(NULL, evlist->mmap_len, prot, 260 MAP_SHARED, fd, 0); 261 if (evlist->mmap[cpu].base == MAP_FAILED) { 262 if (evlist->cpus->map[cpu] == -1 && evsel->attr.inherit) 263 ui__warning("Inherit is not allowed on per-task " 264 "events using mmap.\n"); 265 return -1; 266 } 267 268 perf_evlist__add_pollfd(evlist, fd); 269 return 0; 270 } 271 272 /** perf_evlist__mmap - Create per cpu maps to receive events 273 * 274 * @evlist - list of events 275 * @pages - map length in pages 276 * @overwrite - overwrite older events? 277 * 278 * If overwrite is false the user needs to signal event consuption using: 279 * 280 * struct perf_mmap *m = &evlist->mmap[cpu]; 281 * unsigned int head = perf_mmap__read_head(m); 282 * 283 * perf_mmap__write_tail(m, head) 284 * 285 * Using perf_evlist__read_on_cpu does this automatically. 286 */ 287 int perf_evlist__mmap(struct perf_evlist *evlist, int pages, bool overwrite) 288 { 289 unsigned int page_size = sysconf(_SC_PAGE_SIZE); 290 int mask = pages * page_size - 1, cpu; 291 struct perf_evsel *first_evsel, *evsel; 292 const struct cpu_map *cpus = evlist->cpus; 293 const struct thread_map *threads = evlist->threads; 294 int thread, prot = PROT_READ | (overwrite ? 0 : PROT_WRITE); 295 296 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0) 297 return -ENOMEM; 298 299 if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0) 300 return -ENOMEM; 301 302 evlist->overwrite = overwrite; 303 evlist->mmap_len = (pages + 1) * page_size; 304 first_evsel = list_entry(evlist->entries.next, struct perf_evsel, node); 305 306 list_for_each_entry(evsel, &evlist->entries, node) { 307 if ((evsel->attr.read_format & PERF_FORMAT_ID) && 308 evsel->sample_id == NULL && 309 perf_evsel__alloc_id(evsel, cpus->nr, threads->nr) < 0) 310 return -ENOMEM; 311 312 for (cpu = 0; cpu < cpus->nr; cpu++) { 313 for (thread = 0; thread < threads->nr; thread++) { 314 int fd = FD(evsel, cpu, thread); 315 316 if (evsel->idx || thread) { 317 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, 318 FD(first_evsel, cpu, 0)) != 0) 319 goto out_unmap; 320 } else if (__perf_evlist__mmap(evlist, evsel, cpu, 321 prot, mask, fd) < 0) 322 goto out_unmap; 323 324 if ((evsel->attr.read_format & PERF_FORMAT_ID) && 325 perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0) 326 goto out_unmap; 327 } 328 } 329 } 330 331 return 0; 332 333 out_unmap: 334 for (cpu = 0; cpu < cpus->nr; cpu++) { 335 if (evlist->mmap[cpu].base != NULL) { 336 munmap(evlist->mmap[cpu].base, evlist->mmap_len); 337 evlist->mmap[cpu].base = NULL; 338 } 339 } 340 return -1; 341 } 342 343 int perf_evlist__create_maps(struct perf_evlist *evlist, pid_t target_pid, 344 pid_t target_tid, const char *cpu_list) 345 { 346 evlist->threads = thread_map__new(target_pid, target_tid); 347 348 if (evlist->threads == NULL) 349 return -1; 350 351 if (target_tid != -1) 352 evlist->cpus = cpu_map__dummy_new(); 353 else 354 evlist->cpus = cpu_map__new(cpu_list); 355 356 if (evlist->cpus == NULL) 357 goto out_delete_threads; 358 359 return 0; 360 361 out_delete_threads: 362 thread_map__delete(evlist->threads); 363 return -1; 364 } 365 366 void perf_evlist__delete_maps(struct perf_evlist *evlist) 367 { 368 cpu_map__delete(evlist->cpus); 369 thread_map__delete(evlist->threads); 370 evlist->cpus = NULL; 371 evlist->threads = NULL; 372 } 373 374 int perf_evlist__set_filters(struct perf_evlist *evlist) 375 { 376 const struct thread_map *threads = evlist->threads; 377 const struct cpu_map *cpus = evlist->cpus; 378 struct perf_evsel *evsel; 379 char *filter; 380 int thread; 381 int cpu; 382 int err; 383 int fd; 384 385 list_for_each_entry(evsel, &evlist->entries, node) { 386 filter = evsel->filter; 387 if (!filter) 388 continue; 389 for (cpu = 0; cpu < cpus->nr; cpu++) { 390 for (thread = 0; thread < threads->nr; thread++) { 391 fd = FD(evsel, cpu, thread); 392 err = ioctl(fd, PERF_EVENT_IOC_SET_FILTER, filter); 393 if (err) 394 return err; 395 } 396 } 397 } 398 399 return 0; 400 } 401